pNFS/flexfiles: Don't tie up all the rpciod threads in resends

We do not want to have rpciod threads perform recursive calls into the
RPC layer since that can deadlock. In particular, having to wait for
a layoutget can be nasty... We want rather to defer scheduling those
retries until we're in the rpc_release() callback, since that is
called from the nfsiod workqueue.
Signed-off-by: NTrond Myklebust <trond.myklebust@hammerspace.com>

fs/nfs/flexfilelayout/flexfilelayout.c

@@ -1243,17 +1243,18 @@ static int ff_layout_read_done_cb(struct rpc_task *task,
 					   hdr->ds_clp, hdr->lseg,
 					   hdr->pgio_mirror_idx);
 
+	clear_bit(NFS_IOHDR_RESEND_PNFS, &hdr->flags);
+	clear_bit(NFS_IOHDR_RESEND_MDS, &hdr->flags);
 	switch (err) {
 	case -NFS4ERR_RESET_TO_PNFS:
 		if (ff_layout_choose_best_ds_for_read(hdr->lseg,
 					hdr->pgio_mirror_idx + 1,
 					&hdr->pgio_mirror_idx))
 			goto out_eagain;
-		ff_layout_read_record_layoutstats_done(task, hdr);
-		pnfs_read_resend_pnfs(hdr);
+		set_bit(NFS_IOHDR_RESEND_PNFS, &hdr->flags);
 		return task->tk_status;
 	case -NFS4ERR_RESET_TO_MDS:
-		ff_layout_reset_read(hdr);
+		set_bit(NFS_IOHDR_RESEND_MDS, &hdr->flags);
 		return task->tk_status;
 	case -EAGAIN:
 		goto out_eagain;
@@ -1403,6 +1404,10 @@ static void ff_layout_read_release(void *data)
 	struct nfs_pgio_header *hdr = data;
 
 	ff_layout_read_record_layoutstats_done(&hdr->task, hdr);
+	if (test_bit(NFS_IOHDR_RESEND_PNFS, &hdr->flags))
+		pnfs_read_resend_pnfs(hdr);
+	else if (test_bit(NFS_IOHDR_RESEND_MDS, &hdr->flags))
+		ff_layout_reset_read(hdr);
 	pnfs_generic_rw_release(data);
 }
 

include/linux/nfs_xdr.h

@@ -1438,6 +1438,8 @@ enum {
 	NFS_IOHDR_EOF,
 	NFS_IOHDR_REDO,
 	NFS_IOHDR_STAT,
+	NFS_IOHDR_RESEND_PNFS,
+	NFS_IOHDR_RESEND_MDS,
 };
 
 struct nfs_io_completion;